vsearch

A vector search tool powered by Qdrant and ONNX Runtime.

Usage

1. Set up Qdrant server.
2. Create a collection with appropriate vector size and distance metric. BGESmallZHV15 uses 512-dimensional vectors, BGELargeZHV15 uses 1024-dimensional vectors. Example for BGESmallZHV15:

curl -X PUT "http://localhost:6333/collections/cases" \
  -H "Content-Type: application/json" \
  -d '{
    "vectors": {
      "size": 512,
      "distance": "Cosine"
    }
  }'

3. Configure config.toml with your settings.
4. Run the application:

cargo build --release

or with CUDA support so that it can use GPU :

• Download onnxruntime with CUDA support from ONNX Runtime releases
• Unzip the package and set the environment variable ORT_DYLIB_PATH to point to the libonnxruntime.so file in the unzipped folder.
• Build with the cuda feature: cargo build --release -F cuda
• Run with CUDA support: ORT_DYLIB_PATH=/path/to/libonnxruntime.so target/release/vsearch Documentation

Note: Ensure you could connect to the Hugging Face, you may need to set the HTTPS_PROXY and HTTP_PROXY environment variables.